Automatic Adenoid Segmentation and 3D Nasopharyngeal Obstruction Quantification Method Based on CBCT

Abstract

This invention discloses an automatic adenoid segmentation and 3D nasopharyngeal obstruction quantification method based on CBCT, establishing a deep learning prediction model. During model training, CBCT images before and after adenoidectomy are used to obtain adenoid annotation results through registration. The nasopharyngeal airway and adenoid annotation results are input into a SegResNet model to learn morphological changes in the nasopharyngeal airway, and multi-reference atlas weighted fusion is employed to improve segmentation accuracy under small sample conditions. In the application phase, the CBCT image to be treated is input into the model to predict the nasopharyngeal airway after adenoidectomy, and adenoid segmentation results are obtained through Boolean difference. By calculating the ratio of adenoid volume to nasopharyngeal airway volume, the 3D adenoid-nasopharyngeal ratio is obtained as an indicator of obstruction degree, thereby achieving automated processing of adenoid segmentation, volume estimation, and obstruction quantification.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical monitoring and management, specifically to a method for automatic adenoid segmentation and quantification of nasopharyngeal three-dimensional obstruction based on CBCT. Background Technology

[0002] Childhood obstructive sleep apnea (OSA) is a breathing disorder caused by upper airway obstruction that can affect a child's cognitive function, craniofacial development, and daily life. Among the many risk factors for childhood OSA, adenoid hypertrophy is considered one of the most significant. This issue is particularly important for dental professionals, especially orthodontists, as it is closely associated with abnormal craniofacial development. While severe OSA is usually diagnosed earlier in pediatrics, mild to moderate cases are often discovered during routine dental checkups or orthodontic evaluations during the mixed dentition stage. Recent research has increasingly focused on the possibility that orthodontic appliances may help alleviate childhood OSA symptoms and even reduce the degree of adenoid hypertrophy.

[0003] However, a major challenge for oral clinicians and researchers is accurately assessing adenoid volume and the degree of obstruction it causes to the nasopharyngeal airway. Current adenoid assessment primarily relies on lateral nasopharyngeal radiographs, nasal endoscopy, and MRI. Lateral nasopharyngeal radiographs provide only two-dimensional information, making it difficult to comprehensively reflect airway narrowing and often underestimating adenoid volume. While nasal endoscopy can visually assess adenoid hypertrophy, it cannot provide three-dimensional information and relies on the clinical judgment of an ENT specialist. Dentists often lack access to this examination, and it requires a high degree of cooperation from children. MRI offers high soft tissue resolution but is expensive and complex to perform, making it unsuitable for routine screening.

[0004] Cone-beam computed tomography (CBCT) is widely used in the oral and maxillofacial region due to its low radiation dose, high resolution, and short scan time. However, its insufficient soft tissue contrast results in blurred boundaries of adenoids in CBCT images, making automated identification difficult with existing segmentation algorithms, thus limiting its application in adenoid assessment. Clinically, there is also a lack of objective, three-dimensional quantitative indicators based on CBCT to reflect the degree of adenoid obstruction of the nasopharyngeal airway.

[0005] It is worth noting that nasopharyngeal endoscopy is primarily used to assess airway space loss due to adenoid hypertrophy, while the purpose of adenoidectomy is to restore the compressed nasopharyngeal airway. Therefore, when assessing adenoid hypertrophy and its degree of upper airway obstruction, clinicians mainly focus on the volume of adenoids that need to be removed. If the nasopharyngeal airway morphology after adenoidectomy can be predicted using preoperative CBCT scans, then the difference between the preoperative morphology and the predicted postoperative morphology represents the adenoids, thus obtaining their three-dimensional information.

[0006] In view of this, the present invention proposes an automatic adenoid segmentation and nasopharyngeal three-dimensional obstruction quantification method based on CBCT. Summary of the Invention

[0007] The purpose of this invention is to provide an automatic adenoid segmentation and 3D nasopharyngeal obstruction quantification method based on CBCT, which solves the problem that adenoids cannot be directly segmented in CBCT images and their impact on the degree of nasopharyngeal obstruction cannot be quantified.

[0008] In a first aspect, the present invention provides a method for automatic adenoid segmentation and nasopharyngeal three-dimensional obstruction quantification based on CBCT, comprising the following steps:

[0009] Step S101: The airway segmentation results obtained from the preoperative CBCT image of the first image and the second image through standardized upper airway threshold segmentation are labeled as fixed image and moving image, respectively, where:

[0010] The first images include CBCT images of the same patient before and after adenoidectomy, with the time interval between the preoperative and postoperative CBCT images not exceeding 1 year.

[0011] The second image includes CBCT images taken at least 6 months after adenoidectomy; the first and second images do not overlap.

[0012] Step S102: Perform multiple rigid registrations on the preoperative CBCT images and postoperative CBCT images of the same patient in the first image to obtain adenoid annotation results, and input the adenoid annotation results as supervision information into the deep learning prediction model;

[0013] Step S103: Train a deep learning prediction model based on the airway segmentation results and adenoid annotation results. The deep learning prediction model will perform non-rigid registration of the moving image to the fixed image. The adenoid annotation results are included in the loss function as a key constraint. The deep learning prediction model learns the change law of airway morphology before and after adenoidectomy and virtually completes the local airway narrowing or collapse area.

[0014] Step S104: Using the deep learning prediction model, reason about the CBCT image to be treated to obtain the corresponding airway morphology after adenoidectomy. By performing Boolean difference with the airway before treatment, the adenoidectomy is automatically segmented to obtain the automatic segmentation result of the adenoidectomy.

[0015] Step S105: Extract the adenoid volume based on the automatic adenoid segmentation results, define and extract the nasopharyngeal airway volume based on the airway segmentation results, and calculate the three-dimensional adenoid-nasopharyngeal ratio using the adenoid volume and the nasopharyngeal airway volume to characterize the degree of nasopharyngeal airway obstruction.

[0016] As a preferred embodiment of the present invention, the processing logic for the standardized upper airway threshold segmentation is as follows:

[0017] When performing segmentation, the spatial range of the upper airway is defined and trimmed: the posterior nasal spine is used as the anterior boundary in the coronal plane, and the soft palate plane is used as the lower boundary in the sagittal plane to ensure that the segmentation area focuses on the key anatomical sites where the adenoids may cause obstruction. The resulting airway mask is used as the airway segmentation result.

[0018] As a preferred embodiment of the present invention, the registration logic between the preoperative CBCT image and the postoperative CBCT image in the first image is as follows:

[0019] First, based on the bony features of the skull base, the preoperative and postoperative CBCT images are globally aligned, and the postoperative CBCT images are transformed into the coordinate system of the preoperative CBCT images.

[0020] Local voxel registration was performed on the non-adenoid region of the posterior nasopharyngeal wall to maximize the anatomical consistency of the non-hypertrophic part. The spatial difference between the registration of preoperative and postoperative CBCT images was defined as the adenoid volume.

[0021] The differential spatial region is labeled to obtain the labeling results of the adenoids; the airway segmentation results obtained from the first image in step S101 and the adenoids labeling results extracted after processing in S102 together constitute the training data of the deep learning prediction model.

[0022] As a preferred embodiment of the present invention, the construction logic of the deep learning prediction model is as follows:

[0023] Network structure: The SegResNet network, which combines the U-Net structure with ResNet residual units, is used for airway feature extraction and airway morphology prediction.

[0024] Loss function: During training, the loss function includes a similarity constraint term based on the Dice similarity coefficient and a deformation field smoothness regularization term; among them, the similarity constraint term is used to guide the moving image and the fixed image to achieve better structural registration, and the smoothness regularization term is used to limit the excessive deformation of the dense deformation field.

[0025] Training data: Fixed images and moving images are randomly paired and used as network input;

[0026] Training objective: By optimizing network parameters through a loss function, the deep learning network is trained on randomly paired fixed and moving images to learn the dense deformation field of airway morphological changes before and after adenoidectomy. Based on the dense deformation field, the morphological change patterns of preoperative and postoperative airway segmentation are characterized.

[0027] As a preferred embodiment of the present invention, the SegResNet network includes:

[0028] The initial convolutional layer maps the input features to 64 channels;

[0029] The encoder path contains four downsampling stages, each with 1, 2, 2, and 4 residual blocks respectively. After each downsampling, the spatial resolution is halved and the number of channels is doubled. The decoder path contains three upsampling stages. Each stage restores the resolution and halves the number of channels through transposed convolution, and then concatenates the results with the corresponding features of the encoder.

[0030] The output layer maps the decoder output to three channels, including a dual-channel input and a dense deformation field output channel.

[0031] As a preferred embodiment of the present invention, the logic for obtaining the automatic adenoid segmentation result is as follows:

[0032] The input CBCT image to be treated and multiple moving images are registered in a trained deep learning network to obtain multiple prediction results. The prediction results are dense deformation fields and corresponding predicted airway morphology after adenoidectomy.

[0033] The average or similarity-based weighted fusion of multiple predicted dense deformation fields with the airway morphology after adenoidectomy is performed to generate the final predicted nasopharyngeal airway morphology after adenoidectomy.

[0034] Boolean difference is performed between the predicted postoperative airway morphology and the upper airway segmentation results obtained from the CBCT images to be treated, and the difference spatial region between the two is extracted. This difference spatial region is defined as the adenoid volume region, and the output is the automatic adenoid segmentation result.

[0035] As a preferred embodiment of the present invention, the logic for obtaining the blocking degree assessment result is as follows:

[0036] The nasopharyngeal airway volume range is defined as follows: the anterior boundary is the coronal plane where the posterior nasal spine is located, the posterior boundary is the posterior wall of the nasopharyngeal airway, the superior boundary is the top of the nasopharyngeal airway, and the inferior boundary is the horizontal plane where the lowest edge of the first cervical vertebra is located; the calculation formula for the three-dimensional adenoid-nasopharyngeal ratio (3D-AN) is as follows:

[0037] ;

[0038] in, This refers to the volume of the adenoids. This refers to the volume of the nasopharyngeal airway.

[0039] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0040] This invention proposes an indirect prediction method based on neural networks to overcome the limitations of cone-beam computed tomography (CBCT) in soft tissue imaging. Based on preoperative CBCT data, this method constructs a deep learning prediction model of the hypothetical upper airway morphology after adenoidectomy. By comparing the differences in upper airway morphology before and after prediction, it achieves automatic identification, segmentation, and volume estimation of the adenoids. Therefore, the upper airway morphology after adenoidectomy can be predicted using CBCT data before treatment, and the automatic adenoid segmentation results can be obtained through Boolean difference.

[0041] By learning the morphological deformation relationship before and after adenoidectomy using a deep registration model, the nasopharyngeal morphology after adenoidectomy can be simulated using pre-treatment images, and then adenoidectomy segmentation can be obtained through Boolean difference. Based on this method, a three-dimensional adenoidectomy-nasopharyngeal ratio (3D-AN) is proposed. By calculating the volume ratio between the adenoidectomy and the nasopharyngeal airway, three-dimensional quantification of nasopharyngeal airway structural obstruction is achieved for the first time. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0043] Figure 1 This is a flowchart of the method for automatic adenoid segmentation and 3D nasopharyngeal obstruction quantification of the present invention;

[0044] Figure 2 This is a schematic diagram of the registration network structure based on the deep learning prediction model of the present invention;

[0045] Figure 3 This is a schematic diagram of the automatic adenoid segmentation results and 3D reconstruction of the nasopharyngeal airway output by the deep learning prediction model of this invention;

[0046] Figure 4 This is a graph showing the relationship between the training set size and the Dice similarity coefficient for performance testing of the model in this invention.

[0047] Figure 5 This is a schematic diagram illustrating the three-dimensional AN ratio (3D-AN) for defining the boundary between the nasopharyngeal airway and adenoids in this invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings.

[0049] Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. The described embodiments are only a part of the embodiments of this application, not all of them. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0050] Example 1

[0051] Please see Figure 1-2 As shown, this embodiment provides an automatic adenoid segmentation and nasopharyngeal three-dimensional obstruction quantification method based on CBCT. It utilizes a deep neural network to predict and register pre-treatment CBCT images, thereby indirectly estimating adenoid volume. This method is suitable for assessing nasopharyngeal obstruction caused by adenoid hypertrophy. The method includes the following steps:

[0052] Step S101: The airway segmentation results obtained from the preoperative CBCT image of the first image and the second image through standardized upper airway threshold segmentation are labeled as fixed image and moving image, respectively, where:

[0053] The first image includes CBCT images of the same patient before and after adenoidectomy, with the time interval between the preoperative and postoperative CBCT images not exceeding 1 year; the airway segmentation results obtained from the preoperative CBCT images are marked as fixed images.

[0054] The second image includes CBCT images of different patients at least 6 months after adenoidectomy; the first and second images do not overlap; the airway segmentation results obtained by thresholding all the second images are labeled as moving images:

[0055] Specifically, the processing logic for the standardized upper airway threshold segmentation is as follows:

[0056] When performing segmentation, the spatial range of the upper airway is defined and trimmed: the posterior nasal spine is used as the anterior boundary in the coronal plane, and the soft palate plane is used as the lower boundary in the sagittal plane to ensure that the segmentation area focuses on the key anatomical sites where the adenoids may cause obstruction. The resulting airway mask is used as the airway segmentation result.

[0057] To further clarify, during the standardized airway segmentation process, anatomical prior conditions can be introduced. For example, the average height and width of the nasopharynx obtained from population statistics can be used to constrain the segmentation results to prevent over- or under-segmentation. Simultaneously, multi-plane (coronal, sagittal, and transverse) reconstruction methods can be used to verify the segmentation results, ensuring that the preserved segmented area completely covers the anatomical sites where adenoids may cause obstruction.

[0058] More specifically, the first images include preoperative CBCT images and postoperative CBCT images of adenoidectomy, with the time interval between the preoperative CBCT images and the postoperative CBCT images not exceeding 1 year, totaling 126 CBCT images from 63 patients.

[0059] The second set of images includes CBCT images taken at least 6 months after adenoidectomy, totaling 25 CBCT images;

[0060] The first and second images are non-overlapping, serving to expand the training samples for the deep learning prediction model. Specifically, by introducing CBCT images of different patients at least 6 months post-surgery, and using pre-operative CBCT images as reference pairs, the deformation diversity during training is increased. The moving images are registered with the fixed images after dense deformation field transformation, and adenoid annotations based on pre- / post-operative registration are introduced into the loss function as a key constraint. This enables the neural network to learn airway and adenoid structural differences across individuals and time periods, thereby improving the model's generalization ability in real clinical scenarios.

[0061] Step S102: Perform multiple rigid registrations on the preoperative and postoperative CBCT images of the same patient in the first image to obtain adenoid annotation results. Use these adenoid annotation results as supervisory information input into the deep learning prediction model. The adenoid annotation results correspond to the actual adenoid structure of the patient in reality, i.e., the true adenoid value.

[0062] Specifically, the registration logic between the preoperative CBCT image and the postoperative CBCT image of the first image is as follows:

[0063] First, based on the bony features of the skull base, the preoperative CBCT images and postoperative CBCT images are globally aligned. The postoperative CBCT images are then transformed into the coordinate system of the preoperative CBCT images, so that the two are highly consistent in the stable areas of the skull base and posterior nasopharyngeal wall.

[0064] Local voxel registration was performed on the non-adenoid region of the posterior nasopharyngeal wall to maximize the anatomical consistency of the non-mastoid part;

[0065] The difference in the registration area between preoperative and postoperative CBCT images in the adenoid hypertrophy region is defined as the adenoid volume. The difference in the spatial region is then labeled to obtain the adenoid annotation results.

[0066] The first image, processed in step S101 to obtain the preoperative airway segmentation result, and the adenoid annotation result extracted after processing in step S102, together constitute the training data for the deep learning prediction model. The adenoid annotation result is introduced as a key supervisory signal into the loss function corresponding to the deep learning prediction model. The adenoid annotation result provides the deep learning prediction model with anatomical prior knowledge of the target region, enabling the model to not only rely on the morphological constraints of airway segmentation but also combine adenoid structural information for joint optimization. Using the adenoid annotation result as a key constraint in the loss function effectively avoids misalignment of adenoid boundaries and volume estimation errors during registration, ensuring the accuracy of subsequent volume calculation and obstruction quantification.

[0067] Step S103: A deep learning prediction model is trained based on the airway segmentation results and adenoid annotation results. This model performs non-rigid registration of moving images to fixed images, and the adenoid annotation results are included in the loss function as a key constraint. The deep learning prediction model learns the changes in airway morphology before and after adenoidectomy and virtually completes local airway narrowing or collapse areas.

[0068] Specifically, the construction logic of the deep learning prediction model is as follows:

[0069] Network structure: The neural network structure is based on SegResNet, which combines U-Net structure and ResNet residual units. The network consists of encoder path and decoder path. The encoder extracts multi-scale features through multi-level residual blocks, and the decoder restores spatial resolution through upsampling and skip connections.

[0070] Loss Function: During training, a similarity constraint term based on the Dice similarity coefficient and a deformation field smoothness regularization term are introduced. The adenoid annotation results are incorporated into the loss function as a key constraint, enabling the deep learning prediction model to more accurately learn the changes in airway morphology after adenoidectomy, especially for local airway narrowing or collapse areas caused by adenoid hypertrophy, achieving virtual completion of preoperative narrowing or collapse areas. The smoothness regularization term limits excessive deformation of the dense deformation field, avoiding discontinuities or excessive wrinkles, thereby ensuring the rationality and stability of the deformation results.

[0071] Training data: During training, fixed images are randomly paired with multiple moving images and input into the neural network; this effectively expands the training set size under small sample conditions, forming a large-scale paired sample, thereby improving the model's generalization ability.

[0072] Training objective: By optimizing network parameters through a loss function, the deep learning network is trained on randomly paired fixed and moving images to learn the dense deformation field of airway segmentation morphological changes before and after adenoidectomy. Based on the dense deformation field, the morphological change patterns of preoperative and postoperative airway segmentation are characterized.

[0073] More specifically, such as Figure 2 The diagram shown is a neural network structure based on SegResNet. It is based on the classic U-Net structure and introduces residual connections of ResNet residual blocks to solve the gradient vanishing problem in deep networks and improve feature extraction capabilities. The input features are mapped to 64 channels through an initial convolutional layer.

[0074] The encoder path comprises four downsampling stages, each containing 1, 2, 2, and 4 residual blocks (ResBlocks) sequentially. After each downsampling, the spatial resolution is halved, while the number of channels doubles (gradually increasing from 64 to 512). Each ResNet residual block typically contains two 3x3x3 convolutions, instance normalization (InstanceNorm3d), ReLU activation, and applies Dropout regularization with a probability of 0.3 to reduce overfitting. It also contains crucial identity skip connections, through which feature maps generated at each stage of the encoder are passed to the corresponding layers of the decoder.

[0075] The decoder path consists of three upsampling stages. In each stage, the spatial resolution is doubled and the number of channels is halved (gradually reduced from 512 to 64) by transposed convolution (ConvTranspose3d). The result is then concatenated and fused with the corresponding skip connection features from the encoder. Finally, the fused features are processed by a residual block.

[0076] The network's final layer maps the 64-channel features output from the decoder into three output channels: the two input channels and the dense deformation field (DDF) for each voxel displacement vector generated when the moving image is registered to the fixed image. To improve model performance and alleviate overfitting, we will perform data augmentation on the training set.

[0077] To further explain, the training logic of the deep learning prediction model is as follows:

[0078] The neural network non-rigidly registers the moving image onto a fixed image. During this process, the neural network trains and learns from the deformation field of the moving image. In the neural network, the airway segmentation results and adenoid annotation results obtained in steps S101 and S102 are used as part of the loss function to constrain and adjust the training results. The loss function expression for the entire training process is as follows:

[0079]

[0080] in: This represents the loss function of a deep learning prediction model. This represents the fixed image, i.e., the airway segmentation result obtained from the preoperative CBCT image of the first image through step S101. The moving image, i.e., the CBCT image in the second image taken at least 6 months after adenoidectomy, is used to obtain upper airway segmentation and annotation through step S101. Represents a dense deformation field, used to visualize moving images. Deformation to a fixed image ;

[0081] This indicates that the first image represents the airway segmentation result obtained from the preoperative CBCT image through step S101 and the adenoid annotation result obtained through step S102. This represents a moving image after applying a dense deformation field; This indicates that the adenoid annotation results are used as supervisory information in the loss function;

[0082] Dice similarity coefficient is used to measure and The higher the similarity value in overlapping regions, the more consistent the two are. Indicates the degree of inconsistency between the two, used to represent the similarity loss based on the Dice similarity coefficient, and measures... and The similarity in overlapping regions is constrained by Dice loss; that is, the higher the Dice similarity coefficient, the smaller the similarity loss.

[0083] The smoothing term penalizes the smoothness of the deformation field 𝜙 to prevent discontinuities or excessive "wrinkles" in the deformation. 𝜆 represents the weighting coefficient that balances "registration accuracy" and "deformation smoothness".

[0084] During training, a pair of fixed and moving images are randomly selected and input into the neural network. The neural network needs to predict the dense deformation field from the moving image to the fixed image. During each training session, the corresponding adenoid annotations will participate in the loss function to constrain the model's prediction results. The output dense deformation field will be applied to the moving image to generate aligned prediction results.

[0085] The loss function is calculated based on fixed images and adenoid annotations; network parameters are updated via gradient backpropagation. In engineering implementation, the batch size is typically set to 1, and equivalent large-batch training is achieved through gradient accumulation. The optimizer can be Adam, with an initial learning rate of approximately 2e-4. To enhance robustness, 3D data augmentation strategies such as random flipping, rotation, and scaling can be introduced.

[0086] In constructing the loss function, the preoperative airway segmentation result obtained from the first image through step S101 and the adenoid annotation result extracted after processing in S102 are selected; instead of directly using a single airway segmentation image from the CBCT image after adenoidectomy in the first image, the reason is as follows:

[0087] During oral CBCT imaging, consistent head position is required to ensure the relative stability of the skull base. Since the posterior nasopharyngeal wall is adjacent to the skull base, and the skull base is a stable bony structure, the posterior nasopharyngeal wall, except for the depressions caused by adenoid hypertrophy, exhibits high stability in CBCT images taken at different time points. Therefore, obtaining adenoid annotation results through rigid registration in step S102 is feasible. However, the anterior nasopharyngeal wall is not a stable bony structure, but mainly composed of soft tissues such as the soft palate. Even with consistent head position, its morphology may still show significant changes in CBCT images taken at different time points. In the application scenario of this invention, the desired goal is that after inputting pre-treatment airway morphology data, the prediction model can output the post-adenoid resection airway morphology inferred from the pre-treatment airway morphology. The focus is on virtual completion of areas of localized upper airway narrowing or collapse caused by adenoid hypertrophy, while maintaining the morphology of areas in the posterior nasopharyngeal wall other than the depressions caused by adenoid hypertrophy as much as possible. Therefore, if the airway segmentation image from a fixed CBCT image after adenoidectomy is used, the morphology of the anterior nasopharyngeal wall will be additionally constrained in the loss function. However, the morphological changes of the anterior nasopharyngeal wall are not the focus of this invention, so this method was not adopted.

[0088] In the training process of this invention, a multi-reference atlas registration strategy is employed. Specifically, each airway segmentation result obtained from a post-adenoidectomy CBCT image in the second image set via step S101 is used as a moving image, and is paired with each airway segmentation result obtained from a pre-operative CBCT image in the first image set via step S101 as a fixed image. The pairing result is then input into the neural network training module for training. This multi-reference atlas registration strategy can significantly expand the amount of training data under limited sample conditions: in small sample datasets, by combining and pairing multiple moving images with multiple fixed images, the number of training samples can be multiplied, thereby maintaining high segmentation accuracy even with limited medical image data. Considering that medical images are often limited by acquisition conditions, ethical requirements, and clinical resources, and that the preprocessing of image data relies on professional knowledge, making it difficult to obtain standardized samples on a large scale, this strategy is particularly suitable for medical image processing tasks.

[0089] In one specific embodiment, the first image set comprises 126 sets of CBCT images from 63 patients. The training and validation sets are divided as follows: 20% (i.e., 13 patients) of the CBCT images are randomly selected, and the upper airway and adenoid double-annotated data obtained after processing in steps S101 and S102 are used as the validation set; the remaining 50 patients' pre- and post-treatment CBCT images, the airway segmentation results obtained in step S101, and the adenoid annotation results obtained after processing in step S102, are used as fixed images in the training set. Each airway segmentation result obtained in step S101 from the CBCT images after adenoidectomy in the second image set is used as a moving image. Although the training set contains only 50 fixed images and 25 moving images, since the fixed and moving images are randomly paired and input into the neural network during training, 50 × 25 = 1250 pairs of training samples can be formed, achieving a training effect similar to a "large-scale dataset" under small sample conditions.

[0090] Through the above method, the present invention can not only significantly increase the number of samples for model training even when medical image data is limited, but also effectively improve the accuracy of airway segmentation results and adenoid segmentation tasks, thereby providing reliable data support for adenoid segmentation modeling based on CBCT images.

[0091] Step S104: Using the deep learning prediction model, reason about the new CBCT images to be treated to obtain the automatic segmentation results of the adenoids;

[0092] Specifically, the logic for obtaining the automatic adenoid segmentation results is as follows:

[0093] The airway segmentation results obtained from the CBCT images to be treated are used as fixed images input into a trained deep learning network. The network is then registered with the moving images and outputs multiple dense deformation fields and corresponding predicted airway morphology after adenoidectomy.

[0094] Multiple prediction results are averaged and fused to generate the final predicted nasopharyngeal airway morphology after adenoidectomy. This involves averaging or weighting the multiple predicted dense deformation fields with the airway morphology after adenoidectomy to generate the final predicted airway morphology after adenoidectomy.

[0095] Then, Boolean difference is performed between the predicted airway morphology and volume after adenoidectomy and the airway volume before treatment, and the difference is extracted as a three-dimensional volume label of the adenoidectomy to realize automatic segmentation and volume calculation of the adenoidectomy.

[0096] To further explain, during the inference phase, this embodiment continues to employ a multi-reference image registration strategy. The moving image and the airway segmentation result (fixed image) obtained from the input CBCT image to be treated in step S101 are input into the neural network to obtain multiple prediction results. The multiple prediction results are averaged to generate the final predicted postoperative airway morphology.

[0097] Traditional single-atlas registration is susceptible to biases in reference sample selection. Multi-reference registration strategies, by introducing diverse information and averaging multiple independently predicted morphological fields, suppress local errors or unreasonable deformations that may occur with single registration. This significantly improves the robustness and generalization ability of predicting the posterior nasopharyngeal wall morphology. It also reduces the model's dependence on individual reference atlases, minimizing bias in the final adenoid segmentation results.

[0098] After the training network is complete, input a CBCT image of a patient with adenoid hypertrophy to be treated. Using a deep learning prediction model, and averaging the results of multi-reference registration, predict the postoperative airway morphology after adenoid removal. Subtract the input airway data of the patient to be treated using Boolean difference to obtain the automatic adenoid segmentation result. This achieves a fully automated, contactless, and low-cost adenoid segmentation workflow, providing reliable input for subsequent calculation of the three-dimensional adenoid-nasopharyngeal ratio (3D-AN).

[0099] like Figure 3As shown in the figure, the red area represents the airway segmentation result, and the blue area represents the adenoid annotation result. The left figure shows the ground truth of adenoids in the validation set, and the right figure shows the adenoid prediction result output by the model based on the method of this invention. As can be seen from the figures, the predicted results are highly similar to the ground truth in morphology. To quantitatively evaluate the consistency between the model prediction results and the ground truth, this invention uses the Dice Similarity Coefficient (DSC) as an evaluation index, which is defined as follows:

[0100]

[0101] Where A represents the set of pixels (voxels) of the predicted result, B represents the set of pixels (voxels) of the ground truth, |A| and |B| are the number of pixels (voxels) in the sets, and |A|∩B| represents the number of pixels (voxels) in the intersection of the two sets. The Dice similarity coefficient ranges from 0 to 1, with a value closer to 1 indicating a closer consistency between the predicted result and the ground truth.

[0102] like Figure 4 As shown, when the number of training samples increases from 1 to 15, the model's prediction accuracy significantly improves, and the rate of improvement is relatively high. This result verifies the effectiveness of the aforementioned multi-reference map registration strategy; even under small sample conditions, the multi-reference map registration strategy can significantly improve the model's training performance and prediction accuracy. Finally, the Dice similarity coefficient of this invention on the validation set is close to 0.9, indicating that the difference between the model's prediction results and the true values ​​is minimal, demonstrating excellent prediction performance.

[0103] Step S105: Based on the adenoid volume and the nasopharyngeal airway volume, calculate the three-dimensional adenoid-nasopharyngeal ratio (3D-AN) to characterize the degree of nasopharyngeal airway obstruction;

[0104] Specifically, the logic for obtaining the blocking level assessment result is as follows:

[0105] like Figure 5 As shown, the volume range of the nasopharyngeal airway is defined as follows: the anterior boundary is the coronal plane where the posterior nasal spine is located, the posterior boundary is the posterior wall of the nasopharyngeal airway, the superior boundary is the top of the nasopharyngeal airway, and the inferior boundary is the horizontal plane where the lowest edge of the first cervical vertebra is located; the calculation formula for the 3D-AN is:

[0106] ;

[0107] in, This refers to the volume of the adenoids. This invention proposes a 3D-AN method to quantify the degree of obstruction of the upper airway by adenoids for the first time, based on the ratio of adenoid volume to nasopharyngeal airway volume. Compared with existing methods such as 2D lateral nasopharyngeal radiographs which only provide planar information, nasal endoscopy which lacks intuitive quantitative capabilities for the 3D structure of adenoids, and MRI which is costly and unsuitable for routine screening, this invention not only provides a complete 3D structure of adenoids and their spatial relationship with the upper airway, but also accurately calculates the percentage and absolute volume change of airway stenosis caused by adenoid hypertrophy. This achieves non-invasive and automated 3D obstruction assessment using CBCT images. This method overcomes the problems of blurred adenoid boundaries caused by insufficient soft tissue contrast in CBCT images and the difficulty of automatic identification by existing algorithms, significantly improving the accuracy and repeatability of adenoid and nasopharyngeal airway assessment. It provides quantifiable and intuitively verifiable 3D objective indicators for pediatric OSA risk screening, surgical outcome prediction, and orthodontic treatment planning.

[0108] This indicator is the first to achieve three-dimensional quantification of the degree of nasopharyngeal airway obstruction. In clinical applications, a threshold can be set based on this indicator to determine the presence of pathological obstruction: when the 3D-AN ratio is less than the threshold ratio, it is considered that there is no pathological obstruction; when the 3D-AN ratio is greater than or equal to the threshold ratio, it indicates the risk of pathological nasopharyngeal obstruction. Although this invention relies on CBCT images before and after adenoidectomy to construct the predictive ability of airway segmentation and adenoid annotation results during the model training phase of the deep learning prediction model, as mentioned above, clinicians mainly focus on the volume of adenoids that need to be removed when assessing adenoid hypertrophy and its degree of upper airway obstruction. Therefore, the clinical application value of this patent is far beyond adenoidectomy treatment and has universal significance for all patients with adenoid hypertrophy.

[0109] More specifically, in practical applications, this method inputs pre-treatment CBCT imaging data of patients with adenoid hypertrophy, combines a deep learning model to predict and simulate the upper airway morphology after adenoid removal, and then uses Boolean difference operations between the predicted results and the pre-operative upper airway segmentation results to accurately extract the adenoid volume. Furthermore, by calculating the ratio of adenoid volume to nasopharyngeal airway volume, a three-dimensional adenoid-nasopharyngeal ratio (3D-AN) is obtained. The 3D-AN directly reflects the degree of adenoid obstruction in the nasopharyngeal airway, providing an objective reference for early clinical screening. Especially in cases of different degrees of obstruction, this method can assist doctors in making tiered decisions: for patients with mild obstruction, conservative interventions such as medication or oral appliances can be prioritized; while for patients with severe obstruction, adenoidectomy can be recommended promptly, thus ensuring efficacy while avoiding overtreatment or undertreatment. Therefore, this invention not only improves the early risk identification ability for patients with adenoid hypertrophy but also provides a scientific basis for the development of individualized treatment plans, possessing significant clinical application value.

[0110] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for automatic adenoid segmentation and 3D nasopharyngeal obstruction quantification based on CBCT, characterized in that, Includes the following steps: Step S101: The airway segmentation results obtained from the preoperative CBCT image of the first image and the second image through standardized upper airway threshold segmentation are labeled as fixed image and moving image, respectively, where: The first image includes CBCT images of the same patient before and after adenoidectomy, with the time interval between the preoperative and postoperative CBCT images not exceeding 1 year. The second image includes CBCT images of different patients at least 6 months after adenoidectomy; the first and second images do not overlap. Step S102: Perform multiple rigid registrations on the preoperative CBCT images and postoperative CBCT images of the same patient in the first image to obtain adenoid annotation results, and input the adenoid annotation results as supervision information into the deep learning prediction model; Step S103: Train a deep learning prediction model based on the airway segmentation results and adenoid annotation results. The deep learning prediction model will perform non-rigid registration of the moving image to the fixed image. The adenoid annotation results are included in the loss function as a key constraint. The deep learning prediction model learns the change law of airway morphology before and after adenoidectomy and virtually completes the local airway narrowing or collapse area. Step S104: Using the deep learning prediction model, reason about the CBCT image to be treated to obtain the corresponding airway morphology after adenoidectomy. By performing Boolean difference with the airway before treatment, the adenoidectomy is automatically segmented to obtain the automatic segmentation result of the adenoidectomy. Step S105: Extract the adenoid volume based on the automatic adenoid segmentation results, define and extract the nasopharyngeal airway volume based on the airway segmentation results, and calculate the three-dimensional adenoid-nasopharyngeal ratio by using the adenoid volume and the nasopharyngeal airway volume to characterize the degree of nasopharyngeal airway obstruction.

2. The method for automatic adenoid segmentation and nasopharyngeal three-dimensional obstruction quantification based on CBCT according to claim 1, characterized in that, The processing logic for the standardized upper airway threshold segmentation is as follows: When performing segmentation, the spatial range of the upper airway is defined and trimmed: the posterior nasal spine is used as the anterior boundary in the coronal plane, and the soft palate plane is used as the lower boundary in the sagittal plane to ensure that the segmentation area focuses on the key anatomical sites where the adenoids may cause obstruction. The resulting airway mask is used as the airway segmentation result.

3. The method for automatic adenoid segmentation and nasopharyngeal three-dimensional obstruction quantification based on CBCT according to claim 2, characterized in that, The registration logic between the preoperative CBCT image and the postoperative CBCT image in the first image is as follows: First, based on the bony features of the skull base, the preoperative and postoperative CBCT images are globally aligned, and the postoperative CBCT images are transformed into the coordinate system of the preoperative CBCT images. Local voxel registration was performed on the non-adenoid region of the posterior nasopharyngeal wall to maximize the consistency of the anatomical structure of the non-hypertrophic part. The spatial area of ​​difference between the registration of preoperative CBCT images and postoperative CBCT images is the adenoid. The differential spatial region is labeled to obtain the labeling results of the adenoids; the airway segmentation results obtained from the first image in step S101 and the adenoids labeling results extracted after processing in S102 together constitute the training data of the deep learning prediction model.

4. The method for automatic adenoid segmentation and nasopharyngeal three-dimensional obstruction quantification based on CBCT according to claim 1, characterized in that, The construction logic of the deep learning prediction model is as follows: Network structure: The SegResNet network, which combines the U-Net structure with ResNet residual units, is used for airway feature extraction and airway morphology prediction. Loss function: During training, the loss function includes a similarity constraint term based on the Dice similarity coefficient and a deformation field smoothness regularization term; among them, the similarity constraint term is used to guide the moving image and the fixed image to achieve better structural registration, and the smoothness regularization term is used to limit the excessive deformation of the dense deformation field. Training data: Fixed images and moving images are randomly paired and used as network input; Training objective: By optimizing network parameters through a loss function, the deep learning network is trained on randomly paired fixed and moving images to learn the dense deformation field of airway morphological changes before and after adenoidectomy. Based on the dense deformation field, the morphological change patterns of preoperative and postoperative airway segmentation are characterized.

5. The method for automatic adenoid segmentation and nasopharyngeal three-dimensional obstruction quantification based on CBCT according to claim 4, characterized in that, The SegResNet network includes: The initial convolutional layer maps the input features to 64 channels; The encoder path contains four downsampling stages, each with 1, 2, 2, and 4 residual blocks respectively. After each downsampling, the spatial resolution is halved and the number of channels is doubled. The decoder path contains three upsampling stages. Each stage restores the resolution and halves the number of channels through transposed convolution, and then concatenates the results with the corresponding features of the encoder. The output layer maps the decoder output to three channels, including a dual-channel input and a dense deformation field output channel.

6. The method for automatic adenoid segmentation and nasopharyngeal three-dimensional obstruction quantification based on CBCT according to claim 1, characterized in that, The logic for obtaining the automatic adenoid segmentation results is as follows: The input CBCT image to be treated and multiple moving images are registered in a trained deep learning network to obtain multiple prediction results. The prediction results are dense deformation fields and corresponding predicted airway morphology after adenoidectomy. The average or similarity-based weighted fusion of multiple predicted dense deformation fields with the airway morphology after adenoidectomy is performed to generate the final predicted nasopharyngeal airway morphology after adenoidectomy. Boolean difference is performed between the predicted airway morphology for adenoid removal and the upper airway segmentation results obtained from the CBCT images of the patient to be treated. The difference spatial region between the two is extracted and defined as the adenoid volume region. The result is then output as the automatic adenoid segmentation result.

7. The method for automatic adenoid segmentation and nasopharyngeal three-dimensional obstruction quantification based on CBCT according to claim 1, characterized in that, The logic for obtaining the blocking level assessment result is as follows: The nasopharyngeal airway volume range is defined as follows: the anterior boundary is the coronal plane where the posterior nasal spine is located, the posterior boundary is the posterior wall of the nasopharyngeal airway, the superior boundary is the top of the nasopharyngeal airway, and the inferior boundary is the horizontal plane where the lowest edge of the first cervical vertebra is located; the calculation formula for the three-dimensional adenoid-nasopharyngeal ratio (3D-AN) is as follows: ; in, This refers to the volume of the adenoids. This refers to the volume of the nasopharyngeal airway.

Images